Existing referring Audio-Visual Segmentation (Ref-AVS) benchmarks primarily focus on explicit expressions whose referents can be resolved through direct visual, acoustic, or linguistic correspondence. Consequently, they provide limited coverage of implicit references that require multi-step reasoning over multimodal evidence. To bridge this gap, we introduce **Reasoning Ref-AVS**, a new task that requires models to infer the referred object from audio-visual-language cues before generating its segmentation mask. We further construct **ReAVS**, a comprehensive dataset with manually annotated referring expressions, multimodal cues, object masks, and reasoning chains, covering diverse reasoning spanning diverse reasoning patterns. Evaluations of representative Ref-AVS models on ReAVS reveal substantial performance degradation when direct cross-modal correspondence is insufficient, exposing a clear gap between multimodal recognition and referent reasoning. Motivated by these challenges, we propose **Re2Seg**, an omni-modal framework that performs object-centric evidence encoding, multi-hop reasoning, and reasoning-conditioned mask decoding. Experiments demonstrate that Re2Seg achieves stronger segmentation performance under reasoning-intensive references and provides interpretable object-inference traces. These results highlight the importance of learning to reason before masking for bridging the gap between multimodal recognition and referring segmentation.

Controllable text-to-speech (TTS) aims to generate speech from text with precise control over prosodic and speaker-related attributes such as pitch, age, and accent. Existing methods mainly rely on natural language prompts or reference audio cloning. However, prompt-based approaches often suffer from the cross-modal semantic gap between textual descriptions and target speech attributes, resulting in imprecise and coarse-grained control, while cloning-based methods are constrained by reference audio and offer limited flexibility. To address these limitations, we propose TTS-Hub, a controllable TTS framework that leverages modular Low-Rank Adaptation (LoRA) components and their arithmetic composition for fine-grained and flexible speech control. Specifically, we first build a comprehensive Data Hub covering 6 high-level categories and 32 fine-grained speech attributes, and use it to construct a LoRA Hub in which each LoRA is specialized for a specific attribute. During inference, TTS-Hub selects and composes the required LoRA modules through simple arithmetic operations, enabling flexible and extensible attribute control without retraining the backbone model. Experiments on two mainstream TTS frameworks demonstrate that TTS-Hub consistently outperforms strong baselines and enables precise and flexible control.

How an economic shock propagates through the heterogeneous decisions of diverse economic roles and emerges as aggregate fluctuations is a central question in economics. Traditional agent-based models hard-code behavioral rules and cannot read language, while recent LLM-agent simulations often omit explicit belief formation, endogenous price formation, or mechanism-grounded state transitions, and rarely let a human step into the loop. We present WonderEcon, a configurable, LLM-driven yet economic-rule-grounded multi-agent simulation system in which households, firms, commercial banks, a central bank, and a government interact across five markets, forming a self-consistent macroeconomic loop. Every agent decides through a four-step natural-language pipeline, reading real news alongside structured data, with every step observable. A player mode, with dedicated scene and character modeling, lets a human take over one household agent, yielding a counterfactual comparison and the Economic Personality Type Indicator (EPTI). On an interest-rate hiking cycle, WonderEcon demonstrates directional plausibility with respect to six well-known economic principles across four LLMs, and 33 participants find the player system satisfying. The engine and interfaces are publicly available.

Exploring effective collaboration among multiple large language models (LLMs) represents an active research direction, with multiagent debate (MAD) emerging as a popular approach. MAD involves LLMs independently generating responses and refining their own responses by incorporating feedback from other agents in a debate manner. However,empirical experiments reveal the suboptimal performance of MAD in complex reasoning scenarios. We attribute this to the potential misleading caused by peer agents with limited individual capabilities. To address this, we propose a novel sequential collaboration framework named Explain-Analyze-Generate(EAG). By decomposing complex tasks into essential subtasks and employing a pipeline approach, EAG enable agents provide constructive assistance to peers, ultimately yielding higher performance. We conduct experiments on the comprehensive complex language reasoning benchmark: BIG-Bench-Hard (BBH). Our method achieves the highest performance on 19 out of 23 tasks, with an average improvement of 8% across all tasks, and incurs lower costs compared to MAD, demonstrating its effectiveness and efficiency.